tekle-eyesus/MediScan

A hybrid offline/online medical AI assistant for Pneumonia detection featuring explainable heatmaps (Grad-CAM), Amharic localization, and secure edge inference. # MediScan ## Overview MediScan is a medical AI diagnostic assistant designed to aid radiologists and general practitioners. The application bridges the gap in specialist availability by providing instant, explainable detection of Pneumonia in chest X-rays. The system utilizes a hybrid architecture: a secure Python backend running a fine-tuned ResNet-50 model for heavy inference and Explainable AI (Grad-CAM), paired with a lightweight Flutter mobile application for point-of-care use. ## Project Demo Click the image below to watch the full demonstration on YouTube: > *Note: This video demonstrates the live inference pipeline, heatmap generation, and medical reporting workflow.* ## Key Features * **AI-Powered Diagnosis:** Utilizes a ResNet-50 Convolutional Neural Network trained on chest X-rays to detect Pneumonia with high confidence. * **Explainable AI (XAI):** Generates Visual Heatmaps (Grad-CAM) to overlay on X-rays, allowing doctors to verify which lung regions triggered the AI's prediction. * **Medical Reporting:** automatically generates professional PDF reports containing the original X-ray, AI heatmap, patient ID, and doctor's details. * **Patient History:** Implements a local SQLite database to track and manage recent patient diagnoses. * **Configurable Settings:** Allows customization of doctor profiles and hospital details for personalized reporting. ## Technology Stack ### Backend (The Intelligence Hub) * **Framework:** FastAPI (Python 3.10+) * **ML Core:** PyTorch, Torchvision * **Image Processing:** OpenCV, Numpy * **Database:** SQLite with SQLModel (ORM) * **Explainability:** pytorch-grad-cam ### Frontend (Mobile Application) * **Framework:** Flutter (Dart) * **State Management:** Provider * **Networking:** Dio (HTTP Client) * **Visualization:** fl_chart (Data visualization), photo_view (Zoomable images) * **Reporting:** pdf, printing packages ## Architecture The project follows a clean client-server …